CL-6499: Port scout's diligence-brief pipeline into workbench as a workflow

[TheGreatAxios]

Summary

Ports scout's workflows/diligence-brief pipeline into workbench's own
workflows/ catalog as a real, deployable @intx/workflow definition —
workflows/diligence-brief. Registered in packages/workflow-catalog's
WORKFLOW_CATALOG, so it deploys and shows up in the web Routines
picker like every other workflow here; it becomes a routine for free
because workbench workflows always do.

What ported cleanly

• The evidence discipline: ground every claim in a tool result, never
  invent a founder/round/metric/competitor, say "insufficient evidence"
  plainly rather than padding.
• The research order: firm memory first, then the live web — mirrors
  scout's own "tool policy" (memory/knowledge-search before web
  research).
• The finalize-then-approve delivery shape, identical to
  last-30-days-research/pain-point-collateral/collateral-generation.

What got rebound (the real constraint)

Scout's tools compile into scout's own sidecar and cannot be pinned
here (CL-5179). Every tool call is rebound to a workbench-published
package instead:

• web research → @corbits/web-search-tools (web_search, Exa)
• firm memory → @corbits/memory-tools (memory_search)
• artifact persistence → @corbits/artifacts-hub's workflow-artifacts
  surface (create + list-recent only — no search-by-field or
  read-by-id, the same limit the sibling @corbits/scout-agent port
  hit, so this workflow only ever writes one artifact per run)

What was trimmed, and why (be blunt about this vs. the original)

• Single step, not scout's parallel per-section step graph. Scout
  drafts an outline plus one parallel step per section, deterministically
  assembled. This repo's execution host has never wired the action
  primitive that graph assumes (same gap last-30-days-research's own
  header documents). Folding it into seven serial steps would blow the
  timebox and multiply timeout risk, so this port follows code-review's
  single-step shape instead: one agent, one turn, everything in one pass.
• Five sections, not seven — Revenue/Capital/Growth collapse into one
  "Traction & Funding" section; Market & Competition folds into "Product &
  Market". A single-turn brief can't sustain three sections' worth of
  genuinely distinct evidence without padding.
• No scoring. Scout's seven weighted dimensions (SCORE_WEIGHTS,
  dims/dimensionReads) are cut entirely — real product value, and
  a deliberate scope trim, not an oversight. Restoring it is the natural
  next step once the parallel-step machinery lands in this repo.
• No source-artifact intake — scout opens sourceArtifactRefs before
  searching; this port has no equivalent trigger field yet.

Full account in workflows/diligence-brief/README.md's "What was
trimmed" section.

Verified end to end

bun test (package-scoped, every tool call stubbed — no live Exa, no
live inference): 24 tests in the new package (definition shape, JSON
round-trip, and the finalize tool actually persisting a stubbed
artifact and returning persisted: true, vs. an honest isError: true
naming the failure on a broken persist — never a silent empty brief).
packages/workflow-catalog (43 tests) and packages/settings-ui's
connector-pin cross-check (6 tests) both still pass after registering
the new entry. bun run typecheck on the new package passes clean.

What I did NOT run

Per the memory constraint on this task: no bun run check, no
repo-root lint/typecheck, no check:packages. Only package-scoped
bun test/bun run typecheck on the touched packages
(workflows/diligence-brief, packages/workflow-catalog,
packages/settings-ui) plus a scoped bun install to link the new
workspace member. check:tool-package-pins does not exist in this repo
yet — I manually verified the pinned versions
(@corbits/web-search-tools@0.0.3, @corbits/memory-tools@0.0.4)
match each package's own package.json.

Test plan

• bun test in workflows/diligence-brief — 24 pass
• bun run typecheck in workflows/diligence-brief — clean
• bun test test/catalog.test.ts in packages/workflow-catalog — 43 pass
• bun test src/connections-pinned-by.test.ts in packages/settings-ui — 6 pass
• Deploy-time verification that the tool-package pins actually resolve
  (same open gap every sibling workflow's README calls out for CL-5999)